AnbaniPy

Georgian Python toolkit for NLP, Transliteration and more. Partially based on anbani.js.

Install

pip install anbani

Transliteration and NLP utilities need only hjson. PDF / e-book extraction (anbani.nlp.utils.ebook2text) additionally requires PyMuPDF:

pip install anbani[pdf]

Quickstart

Transliteration example:

from anbani.core.converter import convert, interpret

interpret("გამარჯობა", "asomtavruli")

# 'ႢႠႫႠႰႿႭႡႠ'

Georgianisation example:

from anbani.nlp.georgianisation import georgianise

georgianise("gamarjoba - rogor xar - rasa iqm - kaia kata - kai erti")

# 'გამარჯობა - როგორ ხარ - რასა იქმ - კაია კატა - კაი ერთი'

Convert ebooks with qwerty encoding to unicode Mkhedruli:

from anbani.nlp.utils import ebook2text
from anbani.core.converter import classify_text
from anbani.core.converter import convert

text = ebook2text("/home/george/Dev/georgian-text-corpus/sources/mylibrary/raw/files/ჩარლზ დიკენსი - დევიდ კოპერფილდი.pdf")
print(text[:300])

print(classify_text(text))

print(convert(text, "qwerty", "mkhedruli")[:300])

# Carlz dikensi daviT koperfildi Tavi pirveli dabadeba me viqnebi gmiri Cemive sakuTari Tavgadasavlisa Tu sxva...

# latin

# ჩარლზ დიკენსი დავით კოპერფილდი თავი პირველი დაბადება მე ვიქნები გმირი ჩემივე საკუთარი თავგადასავლისა თუ სხვა...

Expand contractions:

from anbani.nlp.contractions import expand_text

text = "ილია ჭავჭავაძე (დ. 8 ნოემბერი, 1837, სოფელი ყვარელი — გ. 12 სექტემბერი, 1907, წიწამური)"

print(text)
print(expand_text(text))

# ილია ჭავჭავაძე (დ. 8 ნოემბერი, 1837, სოფელი ყვარელი — გ. 12 სექტემბერი, 1907, წიწამური)
# ილია ჭავჭავაძე (დაბადება 8 ნოემბერი, 1837, სოფელი ყვარელი — გარდაცვალება 12 სექტემბერი, 1907, წიწამური)

Contract them back:

from anbani.nlp.contractions import contract_text

print(contract_text("მასის ატომური ერთეული და ასე შემდეგ"))

# მ.ა.ე. და ა.შ.

Text statistics & lorem

from anbani import toolkit, lorem

toolkit.frequency("ანბანი")   # per-letter frequencies
toolkit.friedman(text)         # index of coincidence
lorem.sentences(12)            # fake Georgian prose
lorem.names(3)                 # fake full names

Command line

Installing the package also installs an anbani command:

anbani interpret "gamarjoba"                      # -> ᲒᲐᲛᲐᲠᲯᲝᲑᲐ
anbani convert "ანბანი" -f mkhedruli -t asomtavruli
anbani georgianise "gamarjoba"
anbani expand "ვნახოთ ა. შ."
anbani lorem -w 8

Interactive TUI

New in 3.1: a full-screen terminal UI. Launch it with anbani tui (or python -m anbani tui). Zero extra dependencies — it is hand-rolled ANSI, and imports stay lazy so import anbani and the plain CLI are unaffected.

anbani tui

Five tabs, switch with 1–5 or tab / shift+tab:

Tab	What it does
Converter	live transliteration; m toggles auto/manual, f/o pick source/target, s swaps
Alphabet	letter grid + detail card; hjkl to move, c cycles script, a toggles archaic, enter copies
Lorem	fake Georgian text; k kind, +/- amount, g generate
NLP	georgianise / latinise / expand / contract; m mode, b georgianise level (fast/balanced/accurate)
Toolkit	letter-frequency bars + Friedman index of coincidence

Converter — live transliteration with script auto-detection:

Alphabet — grid + per-letter detail across every Georgian script, IPA, braille, numeral:

NLP — georgianise / latinise / expand / contract:

Toolkit — letter-frequency bars and the Friedman index of coincidence:

Global keys: ? help overlay · i/enter edit · y copy (via OSC 52) · q or ctrl+c quit. Minimum terminal size is 60×16.

Notes:

• No Georgian keyboard? Pick from: qwerty in the Converter (type Latin, get Georgian), or paste — bracketed paste is supported in edit mode.
• Copy uses OSC 52; some terminals block it, in which case the copy silently does nothing.
• Not supported on native Windows (needs a POSIX terminal); use WSL or the npm package there.

Parity with anbani.js

anbani.py and anbani.js share one behavior contract, proved in CI by the byte-identical spec/golden.json that both test suites run. As of 3.0 they are feature-equivalent; the only intentional gaps are ebook2text (Python-only, needs PyMuPDF) and the browser/UMD bundle (js-only).

3.0 — breaking changes

• classify_text returns "unknown" for undetectable / mixed-script text and now reports the four bicameral scripts; interpret raises on "unknown".
• convert / interpret raise ValueError on an unsupported source or target (was an assertion, or a silent passthrough for targets).
• georgianise defaults to mode="balanced" (was "accurate").

To-Do

Feel free to fork this repo!

• Tokenizer
• Transliteration
• Expand contractions
• ebook2text converter
• Stemmer
• Lemmatizer
• Stopwords

Resources used

• http://www.nplg.gov.ge/civil/statiebi/wignebi/qartul_enis_marTlwera/qartul_enis_marTlwera_tavi-12.htm
• http://www.nplg.gov.ge/civil/upload/Semokleba.htm